In 2019, the Australian Wildlife Conservancy (AWC), a prominent conservation nonprofit, initiated the development of an artificial intelligence (AI) model designed to identify animals captured in camera-trap images. Despite facing constraints in funding and resources, the dedicated team has made remarkable progress, successfully training the model to recognize 44 distinct species, including native dingoes, kangaroos, woylies, and introduced predators such as cats and foxes.
Recent developments have injected new life into the project, thanks to a substantial funding grant from the Australian government’s Innovative Biodiversity Monitoring Grants Program. The A$750,000 ($492,000) grant is poised to significantly amplify the model’s capabilities, almost tripling its capacity. Damien Kerr, Chief Information and Technology Officer at AWC, expressed in an email interview with Mongabay, “The funding will allow us to employ additional staff to gather the source data, train and test the models, and purchase additional AI-processing power in the cloud.”
Camera traps have long been instrumental in wildlife tracking and monitoring. Despite the emergence of cutting-edge technologies, camera traps remain one of the most widely utilized tools for biodiversity surveys. However, the increased capabilities of these devices have led to a surge in data collection, creating a bottleneck for scientists and researchers attempting to analyze and process the vast amount of images to identify individual species accurately.
Holly Sitters, one of the leads in the AWC project and Senior Wildlife Ecologist, highlighted the challenges, stating, “While species like koalas and echidnas are easily recognizable, it can take a lot of practice to confidently distinguish a bush rat from a swamp rat.”
To address this bottleneck, the AWC team decided to integrate artificial intelligence into their workflow. By automating the image tagging process, the team aims to expedite results generation, providing land managers with timely information. Sitters explained, “By automating the process of tagging images, it will be possible to generate results in a fraction of the time it takes when processing images manually.”
The training of the AI model involves utilizing data from AWC’s extensive library of camera-trap images collected over the years from their network of sanctuaries. For each species, the team employs nearly 1,000 high-quality images to train the model effectively. Sitters elaborated on the training process, stating, “The training images must show the species from a range of angles, under a range of light levels, and ideally in a range of habitats.”
Once trained, the model’s classifications are compared with those of national network experts to measure its accuracy. Adjustments and refinements are made to the model as necessary. Beyond monitoring and studying native species, the AI model also holds promise in identifying harmful invasive species. Sitters explained, “For example, the presence of an introduced animal at a site may trigger prompt management action to prevent harm to native species.”
With the infusion of federal funding, the AWC team aims to expand the model’s capabilities by identifying up to 120 native species. The challenge lies in obtaining thousands of images for each species, particularly for rare and threatened ones. Kerr emphasized that the funding would assist in establishing guidelines and standards to maintain consistency across various sources and testing metrics.
“Species recognizers of very common fauna are commonplace, but models of rare, threatened, or endangered species are much harder to build with high performance, and it is precisely those species that are in need of monitoring,” Kerr stated. The team aspires to release models encompassing such species for conservation purposes, paving the way for future technologists and conservationists to build upon their advancements.
